WizardLM is a fine-tuned 7B LLaMA model. It fine-tunes following the dialogue through a large number of commands of varying difficulty. The novelty of this model is the use of LLM to automatically generate training data.

The WizardLM model uses a new method called Evol-Instruct (a new method to improve the ability of LLM by using LLM on behalf of humans to autonomously batch generate open instructions of various difficulty levels and technical ranges), through 70k computer-generated Instructions for training, the method generates instructions with different levels of difficulty.

Evol-Instruct uses the following five operations to extend hints:

• add constraints

• deepen

• Concrete

• Add inference steps

• complex input

These operations are applied sequentially to the initial command to make it more complex, and the reply is generated by the LLM.